AI-Assisted Work

A leverage framework for AI-assisted work.
AI-assisted work is model capability times tool surface times user fluency. Read for how the three multiply, since the lowest caps the rest.
AI-ASSISTED WORK
= Model capability  ×  Tool surface  ×  User fluency
×
Multiplier 01
Model Capability
Reasoning & generation What the underlying model can produce in isolation
Shell / execution access Whether the agent can actually run commands, not just write them
Composition across steps Chaining actions into a multi-step workflow
×
Multiplier 02
Tool Surface
CLI coverage Does the tool expose an interface an agent can drive directly?
Programmable core CLI & GUI as thin clients over one shared API
↳ why CLI beats a browser agent
① Contract stabilityVersioned commands. GUI flows break on redesigns
② Speed & costSeconds vs minutes. Near-zero vs token-heavy
③ AuditabilityShell history, scripts, CI logs. Diffable and replayable
④ Architectural signalCLI presence implies a programmable core underneath
×
Multiplier 03
User Fluency
Agent fluency Knowing what the AI can do, where it gets cautious, how to push back
Tool fluency Knowing the domain well enough to scope, direct, and verify
Three places fluency works Pushing back · scoping delegation · verifying output
↳ the fluency matrix
Low Tool
High Tool
High
Agent
PartialPushes back but can't specify the test
Full leverageSpecific pushback, calibrated delegation, real verification
Low
Agent
Both cappedAI treated as smarter autocomplete
Under-usedAccepts first cautious answer. Verifies well
⚠ Deliberate exceptions · where friction is the protection
Some capabilities are better without CLI access. The verified, browser-mediated flow is the design. Programmatic access would be a security regression.
Billing & paymentAttaching a card, plan upgrades, financial consequences
Identity verificationAccount creation, signed-in human required
Security-critical actionsService account keys, credentials, secret scope
MFA-gated flowsThe friction is the point of the step
▼ WHAT THIS LOOKS LIKE IN PRACTICE ▼
Division of labour · Firebase deployment
P0

Prerequisites

◆ Human

Install tooling, create accounts, sign in. One-time setup per machine.

P1

Local Config

◆ Agent

Init Firebase, write config, security rules, .gitignore, deploy scripts.

P2

Secrets

◆ Human · Browser

Generate service account key, load GitHub secrets, vault the JSON.

P3

CI Workflows

◆ Agent

Write deploy.yml and preview.yml GitHub Actions workflows.

P4

Verify

◆ Human

Enable auth, run first deploy, confirm Actions succeed end-to-end.